app.py

# login as a privileged user.
import os
HF_TOKEN = os.environ.get("HF_TOKEN")

from huggingface_hub import login
login(token=HF_TOKEN)

from threading import Thread
from typing import Iterator

import gradio as gr
import spaces
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer

from pyreft import ReftModel, get_intervention_locations

MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))

DESCRIPTION = """\
# ReFT-Chat (Llama-2 7B with 1K examples)

### A chatbot built with ReFT and Llama-2 7B. It is trained with 1K training examples from the unpaired [Ultrafeedback dataset](https://huggingface.co/datasets/openbmb/UltraFeedback). It is not good at multi-turn conversations. You can train your own ReFT agent and share it on HuggingFace by following this [tutorial](https://github.com/stanfordnlp/pyreft/tree/main/examples/gradio/train_and_share.ipynb)!

#### This should only be used for research purposes. We did not conduct additional safety training with ReFT. We evaluate this model using [Alpaca-eval](https://github.com/tatsu-lab/alpaca_eval). Performance results can be found in [our ReFT paper](https://arxiv.org/abs/2404.03592). Our model inherits all the underlying risks associated with Llama. See terms outlined below.
"""

LICENSE = """
<p/>

---
As a derivate work of [Llama-2-7b-chat](https://huggingface.co/meta-llama/Llama-2-7b-chat) by Meta,
this demo is governed by the original [license](https://huggingface.co/spaces/huggingface-projects/llama-2-7b-chat/blob/main/LICENSE.txt) and [acceptable use policy](https://huggingface.co/spaces/huggingface-projects/llama-2-7b-chat/blob/main/USE_POLICY.md).
"""

if not torch.cuda.is_available():
    DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU.</p>"


if torch.cuda.is_available():
    model_id = "meta-llama/Llama-2-7b-hf"
    model = AutoModelForCausalLM.from_pretrained(
        model_id, device_map="cuda", torch_dtype=torch.bfloat16
    )
    reft_model = ReftModel.load("pyvene/reft_chat7b_1k", model, from_huggingface_hub=True)
    reft_model.set_device("cuda")
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    tokenizer.use_default_system_prompt = True

prompt_no_input_template = """Below is an instruction that \
describes a task. Write a response that appropriately \
completes the request.

### Instruction:
%s

### Response:
"""

@spaces.GPU
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    max_new_tokens: int = 1024,
    temperature: float = 0.6,
    top_p: float = 0.9,
    top_k: int = 50,
    repetition_penalty: float = 1.2,
) -> Iterator[str]:

    # tokenize and prepare the input
    conversation = []
    for user, assistant in chat_history:
        conversation += [f"user: {user} assistant : {assistant}"]
    conversation += [message]
    conversation = "\n".join(conversation)
    prompt = prompt_no_input_template % conversation
    prompt = tokenizer(prompt, return_tensors="pt").to(model.device)
    input_ids = prompt["input_ids"]
    attention_mask = prompt["attention_mask"]

    if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
        input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
        attention_mask = attention_mask[:, -MAX_INPUT_TOKEN_LENGTH:]
        gr.Warning(f"Trimmed input from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")

    intervention_locations = torch.tensor([get_intervention_locations(
        last_position=input_ids.shape[-1], positions="f5+l5",
        num_interventions=len(reft_model.interventions))]).permute(1, 0, 2).tolist()
    
    streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = {
        "base": {"input_ids": prompt["input_ids"], "attention_mask": prompt["attention_mask"]},
        "unit_locations": {"sources->base": (None, intervention_locations)},
        "intervene_on_prompt": True,
        "streamer": streamer,
        "max_new_tokens": max_new_tokens,
        "eos_token_id": tokenizer.eos_token_id,
        "early_stopping": True,
        "no_repeat_ngram_size": 5,
        "repetition_penalty": repetition_penalty,
        "do_sample": False,
    }

    t = Thread(target=reft_model.generate, kwargs=generate_kwargs)
    t.start()

    outputs = []
    for text in streamer:
        outputs.append(text)
        yield "".join(outputs)


chat_interface = gr.ChatInterface(
    fn=generate,
    additional_inputs=[
        gr.Slider(
            label="Max new tokens",
            minimum=1,
            maximum=MAX_MAX_NEW_TOKENS,
            step=1,
            value=DEFAULT_MAX_NEW_TOKENS,
        ),
        gr.Slider(
            label="Repetition penalty",
            minimum=1.0,
            maximum=2.0,
            step=0.05,
            value=1.1,
        ),
    ],
    stop_btn=None,
    examples=[
        ["Hello there! How are you doing?"],
        ["Can you explain briefly to me what is the Python programming language?"],
        ["Explain the plot of Cinderella in a sentence."],
        ["How many hours does it take a man to eat a Helicopter?"],
        ["Write a 100-word article on 'Benefits of Open-Source in AI research'"],
    ],
)

with gr.Blocks(css="style.css") as demo:
    gr.Markdown(DESCRIPTION)
    gr.DuplicateButton(value="Duplicate Space for private use", elem_id="duplicate-button")
    chat_interface.render()
    gr.Markdown(LICENSE)

if __name__ == "__main__":
    demo.queue(max_size=20).launch()